ICLR 2026 Papers — Page 5
International Conference on Learning Representations · 5356 papers
AVERE: Improving Audiovisual Emotion Reasoning with Preference Optimization
Ashutosh Chaubey (University of Southern California), Mohammad Soleymani (University of Southern California)
ClassificationRecognitionSupervised Fine-TuningReinforcement LearningPrompt EngineeringVideoTextMultimodalityBenchmarkAudio
🎯 What it does: This paper proposes a benchmark specifically designed for evaluating emotional reasoning in multimodal large language models (EmoReAlM) and a training method based on multimodal direct preference optimization (AVEm-DPO), aiming to eliminate irrelevant audio-visual cue associations and audio-visual hallucination issues in models.
AVEX: What Matters for Animal Vocalization Encoding
Marius Miron (Earth Species Project), Matthieu Geist (Earth Species Project)
ClassificationRecognitionRetrievalConvolutional Neural NetworkTransformerSupervised Fine-TuningBenchmarkAudio
🎯 What it does: Conduct a large-scale empirical study on training methods for animal vocalization encoders, evaluating the impact of different models, data mixing strategies, and training protocols on multi-task performance (classification, detection, individual identification, spectrogram discovery), and propose an optimal training workflow.
Avey-B
Devang Acharya (Avey AI), Mohammad Hammoud (Avey AI)
ClassificationRetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Designed and implemented a fully bidirectional encoder Avey-B based on Avey for efficient pre-training and downstream fine-tuning in resource-constrained environments;
AVoCaDO: An Audiovisual Video Captioner Driven by Temporal Orchestration
Xinlong Chen (Kling Team, Kuaishou Technology), Liang Wang (Chinese Academy of Sciences)
GenerationTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityAudio
🎯 What it does: Proposed the AVoCaDO audio-visual describer
Avoid Catastrophic Forgetting with Rank-1 Fisher from Diffusion Models
Zekun Wang (Georgia Institute of Technology), Christopher J. MacLellan
Knowledge DistillationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Investigate the gradient geometry of diffusion models, proposing a continual learning framework that combines rank-1 Fisher-based EWC with generative replay.
AWM: Accurate Weight-Matrix Fingerprint for Large Language Models
Boyi Zeng (Shanghai Jiao Tong University), Zhouhan Lin (Shanghai Jiao Tong University)
Anomaly DetectionTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a training-agnostic, weight matrix-based LLM fingerprinting method that can detect whether a model is based on an existing base model.
BA-LoRA: Bias-Alleviating Low-Rank Adaptation to Mitigate Catastrophic Inheritance in Large Language Models
Yupeng Chang (Jilin University), Yuan Wu (Jilin University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Proposes BA-LoRA, a parameter-efficient fine-tuning method that mitigates 'catastrophic forgetting' during the transfer of large language models through output space regularization.
Back to Square Roots: An Optimal Bound on the Matrix Factorization Error for Multi-Epoch Differentially Private SGD
Nikita Kalinin, Christoph H. Lampert (Institute of Science and Technology)
OptimizationSafty and PrivacyConvolutional Neural NetworkTransformerImageText
🎯 What it does: Proposed a novel matrix factorization method called Banded Inverse Square Root (BISR) for multi-round differential privacy SGD, derived error upper bounds, and proved its asymptotic optimality;
BAH Dataset for Ambivalence/Hesitancy Recognition in Videos for Digital Behavioural Change
Manuela González-González (ETS Montreal), Eric Granger (ETS Montreal)
RecognitionConvolutional Neural NetworkTransformerVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Constructed a multimodal video dataset BAH containing 300 Canadian participants for automatically identifying conflicting/uncertain emotions in health behavior changes.
Balancing the Experts: Unlocking LoRA-MoE for GRPO via Mechanism-Aware Rewards
Changlian Ma (Nanjing University), Limin Wang (Nanjing University)
OptimizationReinforcement LearningMixture of ExpertsTextMultimodality
🎯 What it does: A mechanism-aware reinforcement learning framework named RO-GRPO, which converts internal routing statistics into reward signals, is studied for fine-tuning LoRA-MoE models under Group Relative Policy Optimization (GRPO).
Bandit Learning in Matching Markets Robust to Adversarial Corruptions
Zheshun Wu (Southern University of Science and Technology), Fang Kong (Southern University of Science and Technology)
OptimizationReinforcement LearningTabular
🎯 What it does: Consider the multi-armed bandit learning problem in decentralized bilateral matching markets with adaptive malicious perturbations, and propose robust algorithms for both known and unknown perturbation levels;
Bandits with Single-Peaked Preferences and Limited Resources
Omer Ben-Porat (Technion-Israel Institute of Technology), Rotem Torkan (Technion-Israel Institute of Technology)
OptimizationReinforcement Learning
🎯 What it does: Studied a single-peaked preference (SP) matching problem under budget constraints, proposing an offline optimal matching algorithm SP-MATCHING and online learning algorithms MVM (known SP structure) and EMC (unknown SP structure), and provided their theoretical guarantees and computational complexity.
BANZ-FS: BANZSL Fingerspelling Dataset
Xin Shen (University of Queensland), Xin Yu (Australian Institute for Machine Learning, Adelaide University)
RecognitionObject DetectionPose EstimationConvolutional Neural NetworkVideoTextBenchmark
🎯 What it does: Constructed BANZ-FS, a large-scale, hierarchically annotated dataset covering two-hand finger spelling in Australian, British, and New Zealand Sign Languages (BANZSL), comprising approximately 35,000 video clips from three sources: news broadcasts, laboratory recordings, and online videos;
BAPO: Stabilizing Off-Policy Reinforcement Learning for LLMs via Balanced Policy Optimization with Adaptive Clipping
Zhiheng Xi (Fudan University), Xuanjing Huang (Fudan University)
OptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed and validated the BAPO (Balanced Policy Optimization with Adaptive Clipping) algorithm, specifically addressing gradient instability, sharp entropy decline, and training collapse caused by data obsolescence in large language models during offline reinforcement learning.
BAR: Refactor the Basis of Autoregressive Visual Generation
Zhicong Tang (Tsinghua University), Baining Guo (Microsoft Research Asia)
GenerationTransformerImageTextMultimodality
🎯 What it does: Propose the Basis Autoregressive (BAR) framework, which reprojects token sequences from traditional autoregressive models into a new basis space via a learnable linear transformation matrix A, enabling adaptive optimization for image generation.
BARREL: Boundary-Aware Reasoning for Factual and Reliable LRMs
Junxiao Yang (Tsinghua University), Minlie Huang (Tsinghua University)
Explainability and InterpretabilityTransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposes the BARREL framework to intervene in overthinking issues (Last-Minute Guessing and Second-Thought Spiraling) in large-scale inference models, enhancing their factual reliability and ability to deny unknown information.
Barriers for Learning in an Evolving World: Mathematical Understanding of Loss of Plasticity
Amir Joudaki (ETH Zürich), Fartash Faghri (Apple)
OptimizationRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper formally defines and analyzes the 'loss of plasticity' (LoP) problem in deep networks within non-stationary environments through dynamical systems theory, proposing that LoP can be viewed as gradient descent being trapped in low-dimensional invariant manifolds in the parameter space. It identifies two types of traps: frozen units and cloned units, and proves that gradient optimization cannot escape from them. Meanwhile, it reveals the tension between low-rank feature compression and plasticity loss. Finally, experiments are conducted to verify the theory and explore methods such as normalization, noise injection, and Dropout to escape or prevent LoP.
BaseReward: A Strong Baseline for Multimodal Reward Model
YiFan Zhang, Liang Wang (CASIA)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: Through systematic experimental research, we propose a complete 'recipe' for constructing high-performance multimodal reward models and introduce an efficient baseline model called BaseReward.
Batch Pruning by Activation Stability
Md Mustakin Alam (University of Louisiana at Lafayette), Aminul Islam (University of Louisiana at Lafayette)
Computational EfficiencyConvolutional Neural NetworkTransformerImageText
🎯 What it does: Propose a batch pruning method (B-PAS) based on activation stability, dynamically identifying and removing batches whose learning contributions have approached zero during training to accelerate training.
Battery Fault: A Comprehensive Dataset and Benchmark for Battery Fault Diagnosis
Qingdi Liu (University of Shanghai for Science and Technology), Qiang Li (University of Shanghai for Science and Technology)
Data SynthesisAnomaly DetectionConvolutional Neural NetworkRecurrent Neural NetworkMultimodalityTime SeriesBenchmark
🎯 What it does: Explored a large-scale, real-world vehicle data-generated electric vehicle battery fault diagnosis dataset and benchmark framework
Bayes Adaptive Monte Carlo Tree Search for Offline Model-based Reinforcement Learning
Jiayu Chen (University of Hong Kong), Jeff Schneider (Carnegie Mellon University)
Supervised Fine-TuningReinforcement LearningWorld ModelBenchmarkPhysics Related
🎯 What it does: In offline model-based reinforcement learning, the uncertainty of the world model is modeled as a Bayesian adaptive MDP, and planning is performed using Bayesian adaptive Monte Carlo Tree Search (Continuous BAMCP). Subsequently, an executable policy is distilled from the search results.
Bayesian Attention Mechanism: A Probabilistic Framework for Positional Encoding and Context Length Extrapolation
Arthur S. Bianchessi (Pontificia Universidade Catolica do Rio Grande do Sul), Lucas S. Kupssinskü (Pontificia Universidade Catolica do Rio Grande do Sul)
Representation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the Bayesian Attention Mechanism (BAM), treating positional encoding as a Bayesian prior, and introduce a learnable Generalized Gaussian Prior (GGD-BAM) for long-context inference.
Bayesian Ensemble for Sequential Decision-Making
Rui Liu (JD.com), Jingping Shao (JD.com)
Reinforcement Learning
🎯 What it does: Proposed a Bayesian Ensemble framework that dynamically updates the index distribution of ensemble members in multi-armed contextual bandit and reinforcement learning environments to better model uncertainty and achieve efficient exploration;
Bayesian Evidence-Driven Prototype Evolution for Federated Domain Adaptation
Xiaoyang Yi (Nankai University), Jian Zhang (Nankai University)
Domain AdaptationFederated LearningConvolutional Neural NetworkTransformerContrastive LearningImageBiomedical Data
🎯 What it does: Propose a Bayesian evidence-driven prototype topology evolution framework FedPTE to address feature distribution differences caused by domain shift in federated learning.
Bayesian Influence Functions for Hessian-Free Data Attribution
Philipp Alexander Kreer (Technical University of Munich), Jesse Hoogland (Timaeus)
Explainability and InterpretabilityConvolutional Neural NetworkTransformerImageTextStochastic Differential Equation
🎯 What it does: Proposed a local Bayesian influence function (local BIF) to attribute training data to individual deep network parameters after training, avoiding dependence on the irreversible Hessian matrix.
Bayesian Neural Networks for Functional ANOVA Model
Seokhun Park (Seoul National University), Yongdai Kim (Seoul National University)
Explainability and InterpretabilityComputational EfficiencyImageTabular
🎯 What it does: Propose Bayesian-TPNN—a Bayesian neural network based on Tensor Product Neural Network, used in functional ANOVA models to automatically learn the number of nodes, interaction groups, and weights, enabling efficient estimation of high-order interactions and reducing computational costs.
Bayesian Parameter Shift Rules in Variational Quantum Eigensolvers
Samuele Pedrielli (BIFOLD), Shinichi Nakajima (BIFOLD)
OptimizationPhysics Related
🎯 What it does: Proposes Bayesian parameter shift rules (Bayesian PSR) and gradient confidence interval (GradCoRe) for gradient estimation and adaptive observation cost control in variational quantum eigensolver (VQE).
Bayesian Post Training Enhancement of Regression Models with Calibrated Rankings
Kevin Tirta Wijaya (Max Planck Institute for Informatics), Vahid Babaei (Fraunhofer SCAI)
OptimizationData-Centric LearningDrug DiscoveryLarge Language ModelTabularBiomedical DataAgriculture Related
🎯 What it does: By building upon existing regression models and employing Bayesian inference to integrate the Gaussian likelihood of the regressor with the Bradley-Terry likelihood equipped with temperature calibration and soft gating, prediction accuracy is enhanced without requiring retraining.
Bayesian Robust Cooperative Multi-Agent Reinforcement Learning Against Unknown Adversaries
Kiarash Kazari (KTH Royal Institute of Technology), György Dán (KTH Royal Institute of Technology)
Recurrent Neural NetworkReinforcement Learning
🎯 What it does: Designed and implemented a robust collaborative multi-agent reinforcement learning framework based on Bayesian decision-making (BATPAL), which discretizes continuous attack types into subsets based on severity against baseline policies, and uses external constraint PPO to learn adversarial policies, ultimately achieving an adaptive robust policy against unknown attacks.
Bayesian Test-Time Adaptation via Dirichlet feature projection and GMM-Driven Inference for Motor Imagery EEG Decoding
Huan LUO, Xu Niu (Xi'an Jiaotong University)
ClassificationDomain AdaptationBiomedical Data
🎯 What it does: Developed BTTA-DG, a gradient-free EEG test-time adaptation framework based on Dirichlet feature projection and GMM Bayesian inference, for motor imagery EEG decoding.
Be Careful When Fine-tuning On Open-Source LLMs: Your Fine-tuning Data Could Be Secretly Stolen!
Zhexin Zhang (Tsinghua University), Minlie Huang (Tsinghua University)
Adversarial AttackLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextFinance Related
🎯 What it does: The paper proposes and verifies an attack method that implants backdoors during the post-training phase. Attackers train backdoors on public LLMs, enabling the precise extraction of queries (prompts) from downstream fine-tuned data through black-box queries after subsequent fine-tuning.
BEAT: Visual Backdoor Attacks on VLM-based Embodied Agents via Contrastive Trigger Learning
Qiusi Zhan (University of Illinois Urbana-Champaign), Daniel Kang (University of Illinois Urbana-Champaign)
Adversarial AttackRobotic IntelligenceSupervised Fine-TuningVision Language ModelVision-Language-Action ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes the BEAT framework, which implants a vision-object-triggered backdoor into Vision-Language Model (VLM)-driven embodied agents, enabling the agents to behave normally under normal conditions but automatically execute attacker-specified multi-step malicious strategies upon encountering trigger objects.
BED-LLM: Intelligent Information Gathering with LLMs and Bayesian Experimental Design
Deepro Choudhury (University of Oxford), Tom Rainforth (University of Oxford)
TransformerLarge Language ModelText
🎯 What it does: The BED-LLM method is proposed by combining large language models with sequential Bayesian experimental design (BED), achieving adaptive multi-round information acquisition;
Bee: A High-Quality Corpus and Full-Stack Suite to Unlock Advanced Fully Open MLLMs
Yi Zhang (Tsinghua University), Shi-min Hu
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningMultimodalityChain-of-Thought
🎯 What it does: This study systematically improves the training quality and performance of fully open-source multimodal large language models by constructing the Honey-Data-15M dataset, the HoneyPipe data cleaning pipeline, and the Bee-8B model.
Behavior Learning (BL)
Zhenyao Ma (Xiamen University), Dongxu Li (Xi'an Jiaotong University)
OptimizationExplainability and InterpretabilityComputational EfficiencyScore-based ModelImageTextTabularBenchmark
🎯 What it does: Propose the Behavior Learning (BL) framework, integrating interpretable utility maximization problems (UMP) with deep compositional structures to achieve interpretable and identifiable machine learning models;
Behavioral Embeddings of Programs: A Quasi-Dynamic Approach for Optimization Prediction
Haolin Pan (University of Chinese Academy of Sciences), Yanjun Wu (University of Chinese Academy of Sciences)
OptimizationComputational EfficiencyRepresentation LearningTransformerBenchmark
🎯 What it does: Proposed a quasi-dynamic program embedding method (Behavioral Embeddings), which applies a set of designed optimization probes on program LLVM IR to quantify the program's response to different optimization sequences, forming a behavioral spectrum. Continuous response vectors are discretized using Product Quantization, and a multi-task Transformer model, PQ-BERT, is then used to learn deep syntax, ultimately obtaining program representations applicable for compiler optimization prediction.
Belief-Based Offline Reinforcement Learning for Delay-Robust Policy Optimization
Simon Sinong Zhan (Northwestern University University of Southampton), Qi Zhu (Northwestern University University of Southampton)
OptimizationTransformerReinforcement LearningSequentialBenchmark
🎯 What it does: Investigates how to train reinforcement learning strategies that remain stable in real-world environments with delays using only offline data without delays.
Benchmarking Bias Mitigation Toward Fairness Without Harm from Vision to LVLMs
Xuwei Tan (Ohio State University), Xueru Zhang (Ohio State University)
Data-Centric LearningVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Proposed the NH-Fair benchmark, which evaluates the performance of visual and multimodal models in fairness without harm using a unified experimental protocol.
Benchmarking ECG FMs: A Reality Check Across Clinical Tasks
M A Al-Masud, Nils Strodthoff (Carl von Ossietzky Universitat Oldenburg)
ClassificationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningTime SeriesBiomedical DataElectrocardiogramBenchmark
🎯 What it does: Systematic evaluation of 26 clinical tasks across 12 public databases, comparing the performance of 8 ECG baseline models with two supervised baselines.
Benchmarking Empirical Privacy Protection for Adaptations of Large Language Models
Bartłomiej Marek (Cispa Helmholtz Center For Information Security), Adam Dziedzic (Cispa Helmholtz Center For Information Security)
Safty and PrivacyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Systematic evaluation and benchmarking of the actual privacy leakage risks for large language models (LLMs) adapted under the pretrain-adapt paradigm with differential privacy (DP).
Benchmarking Large Vision-Language Models on Fine-Grained Image Tasks: A Comprehensive Evaluation
Hongtao Yu (Southeast University), Xiu-Shen Wei (Southeast University)
ClassificationRecognitionGenerationRetrievalKnowledge DistillationContrastive LearningImageTextBenchmark
🎯 What it does: Proposed the FG-BMK fine-grained vision-language model evaluation benchmark, containing 1.01 million questions and 280,000 images, covering two evaluation paradigms: human-oriented and machine-oriented;
Benchmarking LLM Tool-Use in the Wild
Peijie Yu (Tencent), feng zhang
Large Language ModelTextBenchmark
🎯 What it does: Proposed a multi-turn, multi-step tool usage evaluation benchmark called WildToolBench based on real user behavior, and systematically assessed the tool usage capabilities of 57 mainstream LLMs.
Benchmarking Open-ended Segmentation
Cristina González (Universidad de los Andes), Pablo Arbelaez
SegmentationLarge Language ModelVision Language ModelContrastive LearningImageMultimodalityBenchmark
🎯 What it does: Proposed a dedicated evaluation protocol for open segmentation and rebaselined existing methods; simultaneously developed the first multimodal large language model OPAL that uses contrastive learning to align visual regions with text descriptions.
Benchmarking Overton Pluralism in LLMs
Elinor Poole-Dayan (Massachusetts Institute of Technology), Michiel A. Bakker (Massachusetts Institute of Technology)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Developed and released OVERTONBENCH, a benchmark to measure the coverage of large language models in Overton pluralism.
Benchmarking Stochastic Approximation Algorithms for Fairness-Constrained Training of Deep Neural Networks
Andrii Kliachkin (Czech Technical University in Prague), Jakub Marecek (Czech Technical University in Prague)
OptimizationTabularBenchmark
🎯 What it does: Propose a fairness-constrained deep learning training benchmark based on US census data, and implement and compare several stochastic approximation algorithms.
Benefits and Limitations of Communication in Multi-Agent Reasoning
Michael Rizvi-Martel (Mila & University of Montreal), Michael Hahn (Saarland University)
Computational EfficiencyTransformerLarge Language ModelAgentic AITextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a theoretical framework to analyze the expressive power of multi-agent reasoning systems, and derives upper and lower bounds for the number of agents, communication volume, and computational depth for three tasks: retrieval, state tracking, and k-hop reasoning. These theories are subsequently validated through experiments with pre-trained large language models (LLMs).
Benefits and Pitfalls of Reinforcement Learning for Language Model Planning: A Theoretical Perspective
Siwei Wang (Microsoft Research Asia), Wei Chen (Microsoft Research Asia)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningGraph
🎯 What it does: Theoretical and experimental analysis of the effectiveness of reinforcement learning in planning tasks for large language models, comparing the learning dynamics and generalization capabilities of supervised fine-tuning (SFT), policy gradient (PG), and Q-learning.
BEP: A Binary Error Propagation Algorithm for Binary Neural Networks Training
Luca Colombo (Politecnico di Milano), Cesare Alippi (Politecnico di Milano)
Computational EfficiencyImageTime SeriesSequential
🎯 What it does: Proposed a full-binary error propagation algorithm (BEP) to enable end-to-end training of multi-layer binary neural networks (MLP and RNN).
Best-of-Infinity: Asymptotic Performance of Test-Time LLM Ensembling
Junpei Komiyama (Mohamed bin Zayed University of Artificial Intelligence), Masafumi Oyamada (NEC Corporation)
OptimizationTransformerLarge Language ModelText
🎯 What it does: Investigated and implemented the limit (best-of-∞) performance of majority voting based on large language models (LLM), and proposed adaptive sampling methods and weighted LLM integration strategies to approach this limit under limited inference budgets.
Best-of-Majority: Minimax-Optimal Strategy for Pass@k Inference Scaling
Qiwei Di (University of California, Los Angeles), Quanquan Gu (University of California, Los Angeles)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelText
🎯 What it does: Proposed and validated an inference strategy called Best‑of‑Majority (BoM) for Pass@k inference extension, along with a theoretically optimal minimax risk analysis.
Best-of-N through the Smoothing Lens: KL Divergence and Regret Analysis
Gholamali Aminian (Alan Turing Institute), Youssef Mroueh (MIT)
OptimizationText
🎯 What it does: This paper investigates the alignment quality when using Soft Best-of-N (SBoN) and traditional Best-of-N (BoN) for inference under a proxy reward model, providing theoretical upper and lower bounds for KL divergence and regret, and analyzing the impact of proxy reward errors on both strategies.
Best-of-three-worlds Analysis for Dueling Bandits with Borda Winner
Zirui Hu, Fang Kong
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes a unified FTRL algorithm applicable to dueling bandits under the Borda winner setting, achieving optimal three-worlds (BoTW) performance in random, corrupted random, and adversarial environments;
Better Bounds for the Distributed Experts Problem
David Woodruff, Samson Zhou (Texas A&M University)
OptimizationBenchmark
🎯 What it does: Studied the distributed expert online learning problem under a coordinator model, proposing a communication-reward trade-off protocol for general ℓp loss, achieving near-optimal expected reward while significantly reducing communication volume.
Better Learning-Augmented Spanning Tree Algorithms via Metric Forest Completion
Nate Veldt (Texas A&M University), Geoffrey Sanders (Lawrence Livermore National Laboratory)
OptimizationImageTextSequential
🎯 What it does: Propose a learning-enhanced multi-representative point Metric Forest Completion (MFC) algorithm that can complete a minimum spanning tree (MST) or its learning-enhanced version in sub-quadratic time within any metric space.
Better Together: Leveraging Unpaired Multimodal Data for Stronger Unimodal Models
Sharut Gupta (MIT CSAIL), Phillip Isola (MIT CSAIL)
ClassificationRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityAudio
🎯 What it does: This paper proposes an unpaired multimodal representation learning framework called UML (Unpaired Multimodal Learner), which enhances the quality of target modality (e.g., image) representations by leveraging unpaired text, audio, and other auxiliary modalities through shared model weights.
Beware Untrusted Simulators -- Reward-Free Backdoor Attacks in Reinforcement Learning
Ethan Rathbun (Northeastern University), Christopher Amato (Northeastern University)
Adversarial AttackReinforcement LearningImage
🎯 What it does: Investigate the threat of malicious simulators to reinforcement learning systems, proposing a backdoor attack that implants triggers and 'deceptive' states without altering the reward function.
Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in LLMs
Mohammad Tavakoli (University of Alberta), J Ross Mitchell
Large Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the BEAM benchmark and the LIGHT framework to evaluate and enhance the memory and reasoning capabilities of large language models (LLMs) in extremely long conversations.
Beyond Accuracy: Are Time Series Foundation Models Well-Calibrated?
Coen Adler (University of California), Padhraic Smyth (University of California)
TransformerTime SeriesBenchmark
🎯 What it does: This study systematically evaluates the probability calibration performance of current mainstream time series foundation models (TSFM) in zero-shot prediction scenarios.
Beyond Aggregation: Guiding Clients in Heterogeneous Federated Learning
Zijian Wang (Renmin University of China), Qiong Zhang (Renmin University of China)
Federated LearningConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: Proposes FedDRM, a new framework in federated learning that shifts the server's role from passive aggregation to active guidance, learning local models on each client while routing new tasks to the most suitable clients based on their features.
Beyond Binary Preferences: A Principled Framework for Reward Modeling with Ordinal Feedback
Amirhossein Afsharrad (Stanford University), Mohammad Ghavamzadeh (Qualcomm AI Research)
Reinforcement Learning from Human FeedbackBenchmark
🎯 What it does: View the training of the reward model as a discrete ordinal regression problem, derive the negative log-likelihood and all-threshold loss using preference data on the Likert scale, and jointly learn the reward function and thresholds.
Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty
Mehul Damani (Massachusetts Institute of Technology), Jacob Andreas (Massachusetts Institute of Technology)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: By incorporating the Brier score into the reinforcement learning reward function, the language model is trained to simultaneously output answers and confidence during reasoning chain generation, achieving dual improvements in accuracy and calibration.
Beyond Classification Accuracy: Neural-MedBench and the Need for Deeper Reasoning Benchmarks
Miao Jing (Beijing University of Posts and Telecommunications), Shangyang Li (Beijing University of Posts and Telecommunications)
TransformerLarge Language ModelVision Language ModelMultimodalityBiomedical DataMagnetic Resonance ImagingElectronic Health RecordsBenchmark
🎯 What it does: Proposed and released Neural-MedBench—a deep evaluation benchmark for neurology clinical reasoning, integrating multi-sequence MRI, electronic health records (EHR), and clinical notes, containing 200 high-difficulty reasoning tasks;
Beyond DAGs: A Latent Partial Causal Model for Multimodal Learning
Yuhang Liu (Responsible AI Research Centre), Javen Qinfeng Shi (Responsible AI Research Centre)
Domain AdaptationRepresentation LearningVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a 'latent partial causal model' that uses undirected coupling variables to describe the generation process of multimodal data, and prove that multimodal contrastive learning (MMCL) can identifiably recover latent causal variables under this model; further translate the theoretical results into a method for obtaining separable representations from pre-trained CLIP, and validate its advantages in tasks such as few-shot learning and domain generalization on real data.
Beyond Distributions: Geometric Action Control for Continuous Reinforcement Learning
Zhihao Lin (University of Glasgow)
Reinforcement LearningBenchmark
🎯 What it does: Propose Geometric Action Control (GAC), which directly generates continuous actions using spherical geometry, replacing the traditional Gaussian distribution;
Beyond English-Centric Training: How Reinforcement Learning Improves Cross-Lingual Reasoning in LLMs
Shulin Huang (Zhejiang University), Yue Zhang (Westlake University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Investigated the performance differences between reinforcement learning (RL) and supervised fine-tuning (SFT) in multilingual reasoning, exploring their cross-lingual generalization capabilities.
Beyond Ensembles: Simulating All-Atom Protein Dynamics in a Learned Latent Space
Aditya Sengar (EPFL), PATRICK BARTH
Protein Structure PredictionRecurrent Neural NetworkGraph Neural NetworkDiffusion modelScore-based ModelTime SeriesBiomedical DataStochastic Differential Equation
🎯 What it does: Proposes the GLDP framework, exploring three methods for propagating dynamics in the latent space based on a pre-trained LD-FPG encoder-decoder, and verifies its long-term stability, thermodynamic, and dynamic performance across multiple protein systems.
Beyond Entity Correlations: Disentangling Event Causal Puzzles in Temporal Knowledge Graphs
Qian Chen (Zhejiang University), Ling Chen (Zhejiang University)
TransformerGraph
🎯 What it does: Propose HEDRA, a temporal knowledge graph event prediction framework that decouples non-causal, pseudo-causal, static causal, and dynamic causal relationships at the event level.
Beyond Fixed: Training-Free Variable-Length Denoising for Diffusion Large Language Models
Jinsong Li (Chinese University of Hong Kong), Dahua Lin (Chinese University of Hong Kong)
GenerationAI Code AssistantTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Proposes a training-agnostic two-stage dynamic length extension method called DAEDAL to address the limitations of fixed-length generation in Diffusion LLM;
Beyond Grid-Locked Voxels: Neural Response Functions for Continuous Brain Encoding
Haomiao Chen (Cornell University), Amy Kuceyeski (Cornell University)
Representation LearningConvolutional Neural NetworkSupervised Fine-TuningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposed an implicit neural response function (NRF) based on coordinates, modeling fMRI brain signals as a continuous function of visual stimuli and MNI spatial coordinates, enabling cross-subject transfer and arbitrary resolution queries.
Beyond Hearing: Learning Task-Agnostic ExG Representations from Earphones via Physiology-Informed Tokenization
Hyungjun Yoon (KAIST), Lili Qiu (University of Texas at Austin)
ClassificationRepresentation LearningTransformerSupervised Fine-TuningAuto EncoderMultimodalityTime SeriesBiomedical Data
🎯 What it does: This paper designs a low-cost in-ear ExG collector called NeuroBuds and collects the DailySense dataset, which includes 50 hours of free-living recordings and 20 hours of recordings covering five-sense tasks; it also proposes the Physiology-informed Multi-band Tokenization (PiMT) scheme, which generates task-agnostic ExG representations through 12 physiology-guided frequency band filtering and patchification, performs self-supervised reconstruction pre-training on free-living data, and finally fine-tunes on multiple tasks.
Beyond In-Domain Detection: SpikeScore for Cross-Domain Hallucination Detection
Yongxin Deng (University of Technology Sydney), Ling Chen (University of Technology Sydney)
RetrievalAnomaly DetectionTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Investigated cross-domain hallucination detection (GHD), proposed SpikeScore based on multi-turn dialogue instability metrics and constructed a threshold detector
Beyond Instance-Level Alignment: Dual-Level Optimal Transport for Audio-Text Retrieval
Wenqi Guo (Shanghai Jiao Tong University), Lei Xu (Shanghai Jiao Tong University)
RetrievalTextMultimodalityAudio
🎯 What it does: Proposed and implemented the DART framework, combining instance-level inverse optimal transport (IOT) and feature-level unbalanced Wasserstein distance for audio-text retrieval alignment.
Beyond Length: Quantifying Long-Range Information for Long-Context LLM Pretraining Data
Haoran Deng (University of California Los Angeles), Yeyun Gong (Microsoft Research)
Data-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Propose the LongFilter framework, which selects data that significantly improves the next token prediction quality in long contexts by measuring the KL divergence between the long and short context prediction distributions for each token, for further pre-training of long context language models.
Beyond Linear Probes: Dynamic Safety Monitoring for Language Models
James Oldfield (Queen Mary University of London), Fazl Barez (University of Oxford)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyText
🎯 What it does: Proposed a Truncated Polynomial Classifier (TPC) for dynamic safety monitoring in the activation space of large language models, supporting variable computational budgets and input-driven cascading evaluations;
Beyond Linear Processing: Dendritic Bilinear Integration in Spiking Neural Networks
Jingyang Ma (Shanghai Jiao Tong University), Douglas Zhou (Shanghai Jiao Tong University)
ClassificationComputational EfficiencySpiking Neural NetworkReinforcement LearningImage
🎯 What it does: Propose a new spiking neuron model DLIF, incorporating a biologically validated two-linear dendritic integration rule, providing theoretical proof and numerical validation.
Beyond Magic Words: Sharpness-Aware Prompt Evolving for Robust Large Language Models with TARE
Guancheng Wan (University of California Los Angeles), Wei Wang (University of California Los Angeles)
OptimizationLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a black-box gradient-free prompt search framework TARE and its adaptive version ATARE based on text sharpness, aiming to find efficient and robust prompts within the semantic neighborhood.
Beyond Magnitude: Leveraging Direction of RLVR Updates for LLM Reasoning
Kexin Huang (University of Science and Technology of China), Jingren Zhou (Independent Researcher)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Investigated the importance of update direction in RLVR and proposed using token-level logp differences (Δlogp) to diagnose and enhance LLM's reasoning capabilities.
Beyond Markovian Drifts: Action-Biased Geometric Walks with Memory for Personalized Summarization
Parthiv Chatterjee (Dhirubhai Ambani University), Tanmoy Chakraborty (IIT Delhi)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: Proposes the Structured Walk framework Walk2Pers, modeling user preference evolution through action-conditioned geometric steps and dual memory channels, and achieving personalized summarization based on this.
Beyond Markovian: Reflective Exploration via Bayes-Adaptive RL for LLM Reasoning
Shenao Zhang (Northwestern University), Yunxuan Li (Google)
Large Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes BARL, a Bayesian reinforcement learning (RL) algorithm for LLM reflective exploration, which can dynamically update beliefs about the environment and proactively reflect to improve reasoning strategies during inference.
Beyond Masks: Efficient, Flexible Diffusion Language Models via Deletion-Insertion Processes
Fangyu Ding (Hong Kong University of Science and Technology), Jiacheng Sun (Huawei Foundation Model Dept)
GenerationComputational EfficiencyDiffusion modelScore-based ModelText
🎯 What it does: Proposed a Deletion-Insertion Diffusion Language Model (DID), achieving more efficient and variable-length language modeling by replacing the masking/recovery mechanism in Masked Diffusion Language Models (MDLM) with a discrete diffusion process involving deletion and insertion.
Beyond Match Maximization and Fairness: Retention-Optimized Two-Sided Matching
Ren Kishimoto (Institute of Science Tokyo), Yuta Saito (Hanuku-kaso, Co., Ltd.)
Recommendation SystemOptimizationGraphTabular
🎯 What it does: This paper proposes and evaluates an algorithm called MRet based on dynamic learning to rank (LTR), aiming to maximize user retention on two-sided matching platforms, rather than solely pursuing the number of matches or fairness.
Beyond Membership: Limitations of Add/Remove Adjacency in Differential Privacy
Gauri Pradhan (University of Helsinki), Antti Honkela (University of Helsinki)
Safty and PrivacyTransformerImageTextTabular
🎯 What it does: Compare and audit the privacy leakage of machine learning models with differential privacy under add/remove and substitute adjacency relations, proposing a method for constructing and auditing suspicious samples (canary) specifically for substitute adjacency relations, and verifying the conservativeness of privacy bounds for add/remove in experiments.
Beyond Multi-Token Prediction: Pretraining LLMs with Future Summaries
Divyat Mahajan (Mila, Universite de Montreal), Kartik Ahuja (FAIR at Meta)
Representation LearningAI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the Future Summary Prediction (FSP) framework, which trains an auxiliary head to predict a compact representation of the future sequence (instead of individual or multiple future tokens), to reduce teacher-forcing issues and enhance long-range reasoning capabilities.
Beyond Noisy-TVs: Noise-Robust Exploration Via Learning Progress Monitoring
Zhibo Hou (University of California, Merced), Wan Du (University of California, Merced)
Reinforcement LearningImage
🎯 What it does: Proposed a noise-robust exploration method based on Learning Progress Monitoring (LPM) and validated its effectiveness in multiple environments.
Beyond Outliers: A Study of Optimizers Under Quantization
Georgios Vlassis (ETH Zurich), Dan Alistarh (ISTA)
OptimizationTransformerLarge Language ModelText
🎯 What it does: Systematically trained and evaluated LLMs with parameter scales ranging from 50M to 1.5B, investigating the performance differences of six optimizers under post-training quantization (PTQ) and quantization-aware training (QAT).
Beyond Pairwise: Empowering LLM Alignment With (Ranked) Choice Modeling
Yuxuan Tang (National University of Singapore), Yifan Feng (National University of Singapore)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes the RCPO framework, unifying LLM alignment with (ranking) selection models by leveraging diverse human feedback such as single best and top-k ranking to achieve more precise alignment through maximum likelihood optimization;
Beyond Pass@ 1: Self-Play with Variational Problem Synthesis Sustains RLVR
Xiao Liang (University Of California), Weizhu Chen (Microsoft)
Data SynthesisTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes an online self-play variant problem synthesis (SVS) strategy, which uses the model itself to generate and solve challenging training sample variants, thereby achieving data augmentation without external labels in RLVR training while maintaining stable policy entropy.
Beyond Penalization: Diffusion-based Out-of-Distribution Detection and Selective Regularization in Offline Reinforcement Learning
Qingjun Wang (Tongji University), Guang Chen (Shanghai Innovation Institute)
Anomaly DetectionReinforcement LearningDiffusion model
🎯 What it does: Proposed DOSER, a framework based on diffusion models for out-of-distribution (OOD) detection and selective regularization in offline reinforcement learning.
Beyond Prompt-Induced Lies: Investigating LLM Deception on Benign Prompts
Zhaomin Wu (National University of Singapore), Bingsheng He (National University of Singapore)
Explainability and InterpretabilityLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a framework called Contact Searching Question (CSQ) to evaluate the spontaneous deception of large language models (LLMs) under non-inductive prompts, and provides two deception metrics based on psychological definitions: deception intent score ρ and deception behavior score δ;
Beyond RAG vs. Long-Context: Learning Distraction-Aware Retrieval for Efficient Knowledge Grounding
Seong-Woong Shim, Byung-Jun Lee (Korea University)
RetrievalTransformerReinforcement LearningTextBenchmark
🎯 What it does: Proposed a learning-based decentralized attention retrieval framework LDAR, which can adaptively select retrieval intervals from similarity distributions to reduce interference with large language models.
Beyond Raw Detection Scores: Markov-Informed Calibration for Boosting Machine-Generated Text Detection
Chenwang Wu (Hong Kong Baptist University), Defu Lian (University of Science and Technology of China)
Anomaly DetectionText
🎯 What it does: Proposes a lightweight calibration method based on Markov random fields, leveraging the proximity similarity and initial instability of token-level detection scores to enhance the effectiveness of detecting machine-generated text.
Beyond Real: Imaginary Extension of Rotary Position Embeddings for Long-Context LLMs
Xiaoran Liu (Fudan University), Xipeng Qiu (Fudan University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Reintroduce the neglected imaginary part information based on RoPE rotational position encoding, constructing dual attention heads that simultaneously include real and imaginary parts, forming RoPE++.
Beyond RLHF and NLHF: Population-Proportional Alignment under an Axiomatic Framework
Kihyun Kim (MIT), Pablo A. Parrilo (MIT)
Recommendation SystemReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningTextTabular
🎯 What it does: Propose a novel preference learning framework that utilizes a feasible population distribution set derived from social choice theory to recover population distributions from pairwise comparison data and construct strategies aligned with population proportions.
Beyond Scattered Acceptance: Fast and Coherent Inference for DLMs via Longest Stable Prefixes
Pengxiang Li (Alibaba Group), Shilin Yan (Alibaba Group)
GenerationComputational EfficiencyTransformerLarge Language ModelDiffusion modelText
🎯 What it does: This paper proposes a novel decoding scheduler called the Longest Stable Prefix (LSP) Scheduler, which achieves efficient and coherent prefix submission in diffusion-based language model (DLM) inference.
Beyond Sequential Reranking: Reranker-Guided Search Improves Reasoning Intensive Retrieval
Haike Xu (Massachusetts Institute of Technology), Tong Chen (University of Washington)
RetrievalLarge Language ModelContrastive LearningTextMultimodalityBenchmark
🎯 What it does: Proposed a new retrieval pipeline called Reranker-Guided-Search (RGS), which improves retrieval accuracy under a limited reranker budget by performing greedy search on document similarity graphs and leveraging reranker preferences to select documents requiring re-ranking.
Beyond Short Steps in Frank-Wolfe Algorithms
David Martínez-Rubio (IMDEA Software Institute), Sebastian Pokutta (Zuse Institute Berlin)
Optimization
🎯 What it does: Proposed a new Frank-Wolfe algorithm (Optimistic Frank-Wolfe) and a class of primal-dual short steps based on primal-dual gaps, along with theoretical convergence proofs and practical experiments.
Beyond Simple Graphs: Neural Multi-Objective Routing on Multigraphs
Filip Rydin (Chalmers University of Technology), Balazs Kulcsar (Chalmers University of Technology)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: Proposed two multi-objective multigraph path planning models based on graph neural networks: GMS-EB, which directly performs autoregressive edge selection on the multigraph, and GMS-DH, which first conducts non-autoregressive pruning followed by autoregressive path construction.
Beyond Skeletons: Learning Animation Directly from Driving Videos with Same2X Training Strategy
Yuan Zeng, Qingmin Liao
GenerationTransformerDiffusion modelVideo
🎯 What it does: This paper proposes DirectAnimator, which generates human portrait animations directly using raw driving video pixels instead of pose estimation;
Beyond Softmax and Entropy: Convergence Rates of Policy Gradients with $\boldsymbol{f}$-SoftArgmax Parameterization $\&$ Coupled Regularization
Safwan Labbi (Centre de Mathématiques Appliquées, French National Center for Scientific Research, École Polytechnique, Institut Polytechnique de Paris), Eric Moulines (Mohamed bin Zayed University of Artificial Intelligence)
OptimizationReinforcement Learning
🎯 What it does: Propose the f-softargmax strategy parameterization coupled with the corresponding f-divergence regularization, derive explicit convergence rates and sample complexity of stochastic policy gradients under no preprocessing conditions, and conduct theoretical analysis and experimental validation on discrete MDPs.
Beyond Spectra: Eigenvector Overlaps in Loss Geometry
Gabriel Mel (Centre de Recerca Matemàtica)
OptimizationExplainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: Proposed a dual-loss local geometric framework, derived general fluctuation and transfer laws, applied to high-dimensional ridge regression and deep networks, and designed a scalable overlapping estimation algorithm.